NVIDIA releases Nemotron-TwoTower diffusion language model

Arxiv CS.CL (NLP/language models)

The 30B-A3B model uses a frozen autoregressive context tower plus a diffusion denoiser tower for iterative generation, as an alternative to traditional autoregressive models. It is built on the Nemotron 3 Nano backbone and released under BF16 precision.

NvidiaNemotronDiffusion

See 1 more source

NVIDIA has released Nemotron-TwoTower-30B-A3B-Base-BF16, an unusual diffusion-based language model built from the Nemotron 3 Nano 30B-A3B backbone.1 day agonikhilprasanthDiscuss

Arxiv CS.CL (NLP/language models)·Fitsum RedaJohn KamaluRoger WaleffeMostofa PatwaryMohammad ShoeybiBryan Catanzaro

15 hours ago